大语言模型训练逻辑启发下的预防医学专业判断能力的培养

Cultivation of Professional Judgment in Preventive Medicine Inspired by the Training Logic of Large Language Models

摘要: 生成式人工智能环境下，预防医学人才培养不仅要解决“学过什么”和“做过什么任务”，还要回答学生如何在持续反馈中形成稳定的专业判断，并在陌生、动态和不完全信息情境中保持这种判断。本文借鉴大语言模型开发中的反馈对齐与泛化评价逻辑，提出预防医学专业判断培养框架。该框架不把反馈理解为一次性纠错，而是将教师、同伴、实践人员和服务对象等多源信息用于校准学生的问题界定、证据选择、风险判断与行动标准，形成“初步判断–多源反馈–证据复核–方案修订–再次决策”的循环；同时通过改变人群、资源、证据和社会条件构造复杂情境迁移任务，检验学生能否在未见情境中重构方案。本文进一步提出以反馈利用质量、方案修订质量和迁移判断表现评价培养效果，并以营养与食品卫生学相关任务作为应用示例。该框架旨在推动预防医学教学由“给出答案后的反馈”转向“通过反馈形成判断”，由熟悉案例表现转向复杂公共卫生情境中的可迁移专业能力。本文将该框架奠基于Kolb的经验学习循环与Vygotsky的最近发展区理论，阐明反馈修订循环如何把学习过程操作化为可观察、可评分的教学活动，并明确界定大语言模型类比的适用边界与人类判断中价值、情感维度的不可替代性；进而提出分阶段、分层次的实施策略：低年级先行引入教师与同伴的双源反馈，高年级逐步纳入实践与社区反馈，并借助在线平台异步收集与管理多源反馈以降低组织成本。

Abstract: In the era of generative artificial intelligence, the training of preventive medicine professionals must address not only what students have learned and which tasks they have performed, but also how they develop stable professional judgment through sustained feedback and maintain it in unfamiliar, dynamic, and information-incomplete situations. Drawing on the logic of feedback alignment and generalization evaluation in large language model development, this paper proposes a framework for cultivating professional judgment in preventive medicine. Instead of treating feedback as one-off error correction, the framework uses multi-source information from teachers, peers, practitioners, and service recipients to calibrate students’ problem definition, evidence selection, risk judgment, and action standards, forming a cycle of “initial judgment—multi-source feedback—evidence re-examination—plan revision—repeated decision-making.” Complex-situation transfer tasks are constructed by varying population, resources, evidence, and social conditions to test whether students can reconstruct their plans in unseen situations. The framework is theoretically grounded in Kolb’s experiential learning cycle and Vygotsky’s zone of proximal development: the feedback-revision cycle operationalizes the experiential learning loop into observable and gradable teaching activities, while the progressive addition of feedback sources and the controlled increase of situational complexity operationalize scaffolding and its gradual withdrawal. The paper further delineates the boundary of the language-model analogy, emphasizing that human professional judgment involves values, emotion, and ethical accountability that model optimization cannot capture. To support implementation, a staged strategy is proposed: lower-year students first receive dual-source feedback from teachers and peers, while practice and community feedback are introduced progressively in higher years; online platforms are used to collect and manage multi-source feedback asynchronously so as to reduce organizational costs. Effectiveness is to be evaluated by the quality of feedback utilization, the quality of plan revision, and transfer judgment performance, with nutrition- and food-hygiene-related tasks as illustrative examples. The framework aims to move preventive medicine teaching from feedback given after answers are produced toward judgment formed through feedback, and from performance on familiar cases toward transferable professional capability in complex public health situations.